Claude Code with QuickSilver Pro: setup, models and cost

Claude Code talks to an Anthropic-compatible API, and it lets you change which server that is. QuickSilver Pro speaks the same API, so Claude Code runs on it unchanged — same tools, same prompts, same models — and you pay QuickSilver Pro prices: $0.50 / $2.50 per million tokens for Sonnet 5.5 instead of Anthropic's $2.00 / $10.00.

You do not need an Anthropic account or a Claude subscription for this. You need a QuickSilver Pro API key from the dashboard. New accounts get $0.05 of free credit, but it only covers the starter models (GPT-6 Luna, Claude Haiku 5.5, DeepSeek V4.1 Flash, Qwen3.8 Flash Next and MiMo-V2.6-Flash), so you can check the setup with --model haiku; Sonnet and Opus need a top-up, from $5.

Setup

Add an env block to ~/.claude/settings.json (create the file if it does not exist):

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.quicksilverpro.io",
    "ANTHROPIC_AUTH_TOKEN": "YOUR_QSP_KEY",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-5-5"
  }
}

• ANTHROPIC_BASE_URL sends every request to QuickSilver Pro. Note there is no /v1 at the end; Claude Code adds it.
• ANTHROPIC_AUTH_TOKEN is your QuickSilver Pro key.
• The three DEFAULT_*_MODEL lines decide what Claude Code's opus, sonnet and haiku aliases mean. Claude Code also uses the Haiku model on its own for small background tasks, so setting it keeps those calls cheap too.

Then start Claude Code as usual and pick a model:

claude --model sonnet

Inside a session, /model switches between the aliases. You can also give a full model ID, such as claude --model claude-fable-5-1.

If you would rather not change your global settings, export the same variables in one shell instead:

export ANTHROPIC_BASE_URL=https://api.quicksilverpro.io
export ANTHROPIC_AUTH_TOKEN=YOUR_QSP_KEY
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-5-5
claude --model claude-sonnet-5-5

Which models to map

Prices are per million tokens. Sonnet 5.5 is the everyday choice; switch to Opus 5.5 with /model opus for the task that needs it, then switch back. The full comparison, including Fable 5.1 and older Claude models, is in Claude API pricing in 2026.

What a task costs

We measured a small one on 2026-10-15: Claude Code with Sonnet 5.5, asked to find and fix a bug in a one-function Python file. It read the file, made the edit and reported back in three API requests.

• Input: about 62,000 tokens in total. Claude Code sends a system prompt and tool definitions of roughly 20,000 tokens with every request, so even a tiny task starts there.
• Output: 225 tokens.
• Cost at Anthropic's list prices: $0.070. On QuickSilver Pro, every leg of that — input, output, cache writes and cache reads — is 75% cheaper.

Most of that input was served from the prompt cache: the first request wrote the 20,000-token prefix to the cache, and the next two read it back at a tenth of the input price. That is why the second and third requests cost about a fifth of the first.

Real sessions are longer — a feature that takes thirty requests with a growing conversation costs proportionally more — but the ratio holds: every token Claude Code sends or receives is billed at the rate in the table above. Each request appears in your dashboard with its model, token counts and cost, so you can see exactly where the money goes.

Things to know

• A notice about auto mode. When Claude Code runs against any server other than Anthropic's, it may print a note that its new auto-mode classifier billing is not available through this connection. Nothing breaks; auto mode keeps working as before.
• Your Claude subscription is not used. With ANTHROPIC_AUTH_TOKEN set, Claude Code bills every request to your QuickSilver Pro balance. Remove the env block to go back to your Anthropic login.
• Usage is prepaid. Requests are deducted from your balance as they happen; there is no monthly invoice. If the balance runs out and auto-recharge is off, requests are refused until you top up.